repoqa: add phi3-medium

Author: ganler

results/repoqa/COMBINED-RESULTS.json

@@ -167,6 +167,90 @@
       }
     }
   },
+  "microsoft/Phi-3-medium-128k-instruct": {
+    "eval_date": "2024-05-22 22:30:02.520421",
+    "train_size": "128k",
+    "scores": {
+      "all": {
+        "0.0": { "pass@1": 0.88 },
+        "0.1": { "pass@1": 0.804 },
+        "0.2": { "pass@1": 0.76 },
+        "0.3": { "pass@1": 0.736 },
+        "0.4": { "pass@1": 0.726 },
+        "0.5": { "pass@1": 0.708 },
+        "0.6": { "pass@1": 0.678 },
+        "0.7": { "pass@1": 0.662 },
+        "0.8": { "pass@1": 0.632 },
+        "0.9": { "pass@1": 0.562 },
+        "1.0": { "pass@1": 0.412 }
+      },
+      "python": {
+        "0.0": { "pass@1": 0.95 },
+        "0.1": { "pass@1": 0.82 },
+        "0.2": { "pass@1": 0.79 },
+        "0.3": { "pass@1": 0.75 },
+        "0.4": { "pass@1": 0.75 },
+        "0.5": { "pass@1": 0.71 },
+        "0.6": { "pass@1": 0.64 },
+        "0.7": { "pass@1": 0.61 },
+        "0.8": { "pass@1": 0.57 },
+        "0.9": { "pass@1": 0.51 },
+        "1.0": { "pass@1": 0.4 }
+      },
+      "cpp": {
+        "0.0": { "pass@1": 0.81 },
+        "0.1": { "pass@1": 0.72 },
+        "0.2": { "pass@1": 0.68 },
+        "0.3": { "pass@1": 0.66 },
+        "0.4": { "pass@1": 0.64 },
+        "0.5": { "pass@1": 0.64 },
+        "0.6": { "pass@1": 0.61 },
+        "0.7": { "pass@1": 0.58 },
+        "0.8": { "pass@1": 0.54 },
+        "0.9": { "pass@1": 0.51 },
+        "1.0": { "pass@1": 0.37 }
+      },
+      "java": {
+        "0.0": { "pass@1": 0.85 },
+        "0.1": { "pass@1": 0.8 },
+        "0.2": { "pass@1": 0.75 },
+        "0.3": { "pass@1": 0.72 },
+        "0.4": { "pass@1": 0.72 },
+        "0.5": { "pass@1": 0.71 },
+        "0.6": { "pass@1": 0.7 },
+        "0.7": { "pass@1": 0.69 },
+        "0.8": { "pass@1": 0.69 },
+        "0.9": { "pass@1": 0.59 },
+        "1.0": { "pass@1": 0.5 }
+      },
+      "typescript": {
+        "0.0": { "pass@1": 0.96 },
+        "0.1": { "pass@1": 0.93 },
+        "0.2": { "pass@1": 0.88 },
+        "0.3": { "pass@1": 0.86 },
+        "0.4": { "pass@1": 0.84 },
+        "0.5": { "pass@1": 0.81 },
+        "0.6": { "pass@1": 0.79 },
+        "0.7": { "pass@1": 0.79 },
+        "0.8": { "pass@1": 0.74 },
+        "0.9": { "pass@1": 0.64 },
+        "1.0": { "pass@1": 0.3 }
+      },
+      "rust": {
+        "0.0": { "pass@1": 0.83 },
+        "0.1": { "pass@1": 0.75 },
+        "0.2": { "pass@1": 0.7 },
+        "0.3": { "pass@1": 0.69 },
+        "0.4": { "pass@1": 0.68 },
+        "0.5": { "pass@1": 0.67 },
+        "0.6": { "pass@1": 0.65 },
+        "0.7": { "pass@1": 0.64 },
+        "0.8": { "pass@1": 0.62 },
+        "0.9": { "pass@1": 0.56 },
+        "1.0": { "pass@1": 0.49 }
+      }
+    }
+  },
   "Qwen/Qwen1.5-32B-Chat": {
     "eval_date": "2024-05-13 01:39:51.588802",
     "train_size": "32k",